AI Can Help Millions, but Only if Governments Scale It Right, Says Kremer

 

World Bank Group Chief Economist Michael Kremer says open access, integration, standards and financing are needed to turn AI pilots into public services.

 

Artificial intelligence can raise farm incomes, improve health care and lift learning in developing countries, but only if governments change how they buy, test and deploy the technology, World Bank Group Chief Economist Michael Kremer said on Thursday (October 15).

 

Speaking at the Leaders' Speaker Series during the IMF-World Bank Annual Meetings, Kremer said the evidence that AI can expand access to public services is already building. The challenge, he argued, is scale.

 

"We're already seeing that governments and firms are using AI to expand access to public services and raise productivity," he said. "The big question is – how do we scale this?"

 

His answer rested on four principles: open-access tools that governments can adapt; integrated systems that combine technology with human expertise; standards and real-world testing to judge quality; and a reliable source of demand, backed by financing, to reward good products.

 

 

A forecast for 50 million farmers

Kremer illustrated the approach with weather forecasting, which he said has reached more than 50 million farmers in India and Ethiopia and, he hoped, would soon reach other countries.

 

Billions of people depend on rain-fed agriculture, he noted, and farmers must time planting to the start of the rainy season. Plant too late and part of the season is lost; plant too early and a brief shower may not be followed by continuous rain, so seeds fail to germinate and fertiliser is wasted. Shifting climate patterns are making the timing harder to judge.

 

 

Randomised trials from several parts of the world consistently show that farmers who receive scientific forecasts change their behaviour, raising incomes and cutting the risk of catastrophic crop losses, he said.

 

A reasonable estimate is that the financial benefits exceed the costs of developing and delivering forecasts by 100 to one or more.

 

AI, he said, is revolutionising forecasting accuracy, lead times and speed, allowing models to run at a fraction of the cost, even on a desktop computer rather than a supercomputer.

 

Many leading models are open access, developed by public-sector bodies and by Indian labs. That openness mattered, Kremer said.

 

India combined several models to produce better predictions than any single one and could blend AI methods with the physics-based models of its own meteorological institutions.

 

Ethiopia has since built a model that incorporates its own data and has piloted it with what he believed were hundreds of thousands of farmers. But the model was not enough.

 

India's forecast was designed around what farmers need to know – whether it is safe to plant – rather than a purely meteorological definition, drawing on agricultural as well as weather expertise.

 

Officials piloted different versions of the message because many farmers did not at first understand probabilistic forecasts. The forecasts were sent by text message to more than 50 million farmers, with voice messages added in some areas as part of a randomised trial.

 

Preliminary results put the effect at an 8.5 per cent increase, Kremer said, adding that delivery and forecast development cost only pennies per farmer. India is now looking to add further forecasts to the system.

 

 

Technology is only part of the cost

The second principle – integration – applies equally to health, he said. He pointed to tools presented earlier in the meetings that allow community health workers to estimate an infant's height and weight from a short video or to screen for tuberculosis from three coughs on a phone.

 

Those tools, he said, accounted for only about 20 per cent of the effort and cost; the other 80 per cent went into working with community health workers and governments to fit them into human and public systems.

 

He also cited a study of a digital app used by up to 100 million community health workers in India, even before AI, which found increases in antenatal care, folic acid supplementation and immunisation for pregnant women.

 

Weather forecasting also required coordination among several ministries and agencies, including agriculture and science, as well as behavioural science and message testing, he said.

 

 

Choosing wheat from chaff

The third principle, standards, addresses what Kremer described as a procurement problem.

 

Public buyers are often told to choose the lowest bid to avoid corruption, but in software, he said, price is only part of the picture: the software is often a small share of a system's overall cost, and a cheap but ineffective education tool wastes teachers' and pupils' time.

 

"Without standards, there's a risk that tools get chosen based on marketing," he said, rather than on quality.

 

He said that in many sectors, low-quality products that look good in demonstrations and make strong claims make it difficult for governments "to tell the wheat from the chaff". This can produce a race to the bottom or lead officials to adopt nothing at all.

 

Kremer said he was part of a group of weather scientists who published guidelines asking developers to state exactly what they are claiming, how it was verified and how problems such as overfitting were addressed.

 

Similar frameworks could be built for education, health and other sectors, tested both in controlled conditions and in real-world settings, and could inform procurement worldwide, possibly through a global effort.

 

He also pointed to innovative procurement in India, where the national government worked with states to link part of vendors' payments to learning gains.

 

Different products were tried in different environments so results could be compared, creating an incentive for quality to improve. In education, he said, some systems have doubled the rate at which students learn and sharply narrowed the gap between low- and high-income countries.

 

 

Creating demand

The fourth principle is demand. If financing is available to buy open-access, integrated, tested systems, Kremer said, the other components can work in harmony.

 

Philanthropy has played that role in weather forecasting to some extent, he said, and international financial institutions could do more. Where a loan supports a sector such as education, health or agriculture and a country wishes to buy software, evidence on the track record of different approaches would help, he said.

 

Clear standards, in turn, would signal to developers, philanthropies and researchers that products that perform well in the real world are likely to be adopted. 

 

Kremer said that, as an economist, he often focuses on commercial motives, but that many people in the private sector and in research are motivated by impact.

 

He said he sees a very bright future and called on institutions to work together on concrete actions to deliver on the promise of AI for development.